Automated Fillet Identification

We propose to develop a fillet identification methodology using visible imagery that can be collected and processed on commodity hardware such as modern smartphones, dramatically increasing the availability of fillet identification technology. In order to accomplish this, we will gather an annotated data set of fillet pictures, traceable to verifiable whole fish images. Using these pictures, we will train an algorithm capable of identifying species from fillet images only, and efficient enough to run on a smartphone. We will further refine this algorithm to be calibrated to proper probabilistic output, including identifying when it is looking at either poor imagery, or a species it has not been trained on.